Publication

Abstract : The main goal of object detection is to identify and find one or more effective targets in still or video data. It covers a wide range of techniques, including image processing, pattern recognition, and machine learning. The scope of this study is to detect small vehicles in an uncontrolled environment from aerial photographs using effective pre-processing and deep learning algorithms. The model architecture is divided into two phases: Training and Detection. Cropping and extraction of the training samples, feature representation, and classification are all part of the training step. The detection phase comprises of extracting the regions of interest, feature extraction, and classification. The tests are carried out using the Vehicle Detection in Aerial Imagery (VEDAI) dataset. In this paper, we explore the feasibility of dimensionality reduction as pre-processing through Principal Component Analysis (PCA) for effective detection of vehicles in aerial imagery; used to remove unwanted features to remove the common misclassifications caused due to ambiguity of small objects. A comparative study between Deep Learning Models such as ResNet50 and MobileNetv1 based on the proficiency to detect small objects, were coupled with PCA pre-processing to provide the observations. ResNet50 gave a classification accuracy of 85%, whereas MobileNetv1 gave a classification accuracy of 76.25%. From the experimental results, when PCA pre-processing is coupled with architectures comprising skip connections like ResNet50, the misclassification rate of vehicles in aerial imagery was brought down drastically, providing a better detection rate comparable to existing benchmark.